Pure on-demand monthly
Expected GPU-hours × on-demand rate × (1 + overhead). Linear in usage.
Calculator
Reserved vs On-Demand Calculator
Compare a GPU reservation to pure on-demand at your expected usage.
Reservations trade a lower hourly rate for a committed monthly bill that you owe whether or not you actually use the GPUs. The calculator shows pure on-demand, reserved-plus-burst, the GPU-hours of stranded capacity if you under-use, and the monthly usage at which the reservation breaks even.
Interactive calculator
Reserved vs on-demand calculator
The accelerator model you are pricing. Selecting one fills in an illustrative starting rate you can edit.
What one GPU costs per hour, from your own quote or provider.
$
The committed hourly rate the provider quotes for the reservation term.
$
How many GPUs the reservation covers, billed every hour of the term whether used or not.
Reservation length in months. Affects the plain-English summary, not the monthly comparison.
months
Total GPU-hours you expect to actually use each month across all the cluster.
h/mo
Extra cost for storage, networking, orchestration, and platform fees, as a percentage of compute.
%
At 1,200 GPU-h/mo, pure on-demand is $10,560/mo vs reserved plus burst at $6,336/mo over a 12-month commit. Reserved plus burst looks cheaper at this usage. The commitment pays for itself with room to spare.
Pure on-demand monthly$10,560
Reserved plus burst monthly$6,336
Stranded GPU-hours per month240 h
Break-even monthly usage720 GPU-h/mo
Usage sensitivity50% → mixed $6,336, pure $5,280 · 100% → mixed $6,336, pure $10,560 · 150% → mixed $9,504, pure $15,840
Starting values are illustrative defaults you can edit — not live ComputeTape benchmark prices. Replace them with a real quote.
How to read the result
What the numbers mean
Reservations are a fixed monthly bill; pure on-demand scales with what you run. The break-even number is the monthly GPU-hours at which the reservation matches the equivalent on-demand spend; above it, reserved-plus-burst is cheaper, below it on-demand wins.
Reserved plus burst monthly
Reserved bill (always paid) plus any extra GPU-hours above committed capacity at the on-demand rate. Above break-even, this beats pure on-demand.
Stranded GPU-hours per month
Committed hours that go unused at your expected usage. Stranded capacity is paid for but produces nothing.
Break-even monthly usage
Reserved hours × reserved rate / on-demand rate, in monthly GPU-hours. The cleanest single number for sizing a commitment against expected demand.
Why variance matters
Average usage is not enough
Two clusters with the same average monthly GPU-hours can land in very different places. A workload that bursts hard then sits idle strands more capacity on a reservation than one that runs at a steady level. The sensitivity row shows what happens at 50%, 100%, and 150% of your expected usage; the wider that spread is in real life, the more cautious the reservation choice should be.
On-demand vs reserved vs spot GPU pricing
Access terms, savings, and interruption risk side by side.
What is GPU utilization?
How utilization is measured and why paid capacity costs more when it sits idle.